You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Checkpoints: human vs. synthetic language supervision on human video (SmolVLA, LIBERO)

Private while the accompanying workshop paper is under double-blind review.

stage2_seed13/{robot_only,synthetic,human,shuffled_human}/   final Stage-2 policies (LeRobot pretrained_model/, step 23500)
stage1_merged/{synthetic,human,shuffled_human}/               lerobot/smolvla_base with the merged Stage-1 VLM (Stage-2 start points)

Base policy: lerobot/smolvla_base@d9f33c94; LeRobot e0d50211.
Stage 1: LoRA (r=16) on SmolVLA's 16-layer VLM, 2,000 steps on 1,934 HoloAssist clips, seed 13; merged, 94-95% of the update survives bf16.
Stage 2: LIBERO (HuggingFaceVLA/libero@86958911), 25% of episodes (427), 23,500 steps, batch 64, VLM frozen, seed 13.
Evaluation as in the paper: libero_spatial + libero_object, init states 0-9 per task, n_action_steps=1, rename image -> camera1, image2 -> camera2. Seed-13 success: robot_only 62.0%, synthetic 59.5%, human 62.5%, shuffled_human 59.5% (200 episodes each).

Run one policy (Linux + GPU, LeRobot with the libero extra):

hf download <this repo> --include "stage2_seed13/human/*" --local-dir ckpt
lerobot-eval --policy.path=ckpt/stage2_seed13/human --env.type=libero --env.task=libero_object \
  --env.task_ids='[0]' --env.init_states=true --eval.n_episodes=10 --eval.batch_size=10 \
  --rename_map='{"observation.images.image": "observation.images.camera1", "observation.images.image2": "observation.images.camera2"}' \
  --policy.n_action_steps=1 --seed=13 --output_dir=outputs/eval

Downloads last month: -; Downloads are not tracked for this model. How to track

Video Preview

Robotics